III: Small: Moving offline learning to rank online, from theory to practice

点击翻译按钮获取中文摘要

Online learning to rank is a modern machine learning technique that adaptively improves result rankings during its interactions with end users. For example, when applied in a search engine system, an online learning to rank solution can estimate the optimal ranking of results by repeating three steps: present a ranked list, collect user feedback (e.g., clicks), then update the ranking for next round of interaction. However, most existing online learning to rank solutions are extended from algorithms originally designed for online optimization, rather than the ranking problem; and thus, their practical performance is often much worse than their offline counterparts. This directly limits their practical acceptance. This project aims to develop a completely new online learning to rank framework, which directly converts the best practices in offline learning to rank online for improved performance and provable guarantees. The key innovation of this project is to break the exponentially large ranking space into quadratic-size pairwise comparisons on the fly, where online learning is only performed on the uncertain pairs of instance rankings. Built on this new online pairwise learning strategy, this project studies multi-objective optimization, collaborative and federated learning to enable online learning to rank in a wider range of application scenarios, such as fair and personalized online learning to rank. The research outcomes, including the developed algorithms, testbeds and evaluation protocols, will be disseminated via an open-source library. The research activities will be incorporated into teaching materials for student training and education in the area of information retrieval and machine learning. The planed outreach to high school students for education about online information techniques, privacy and fairness will increase their awareness of potential risk of privacy breaches and unfairness in online systems.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.